Anthill Inside 2018

On the current state of academic research, practice and development regarding Deep Learning and Artificial Intelligence.

Learning Real-time Object Detection In The Absence of Large-scale Datasets

Submitted by Vijay Gabale (@vijaygabale) on Saturday, 14 April 2018

videocam_off

Technical level

Intermediate

Section

Full talk

Status

Confirmed & Scheduled

View proposal in schedule

Vote on this proposal

Login to vote

Total votes:  +1

Abstract

This talk will focus an area of computer vision, object detection, that involves automatically localising and classifying different objects from photos and videos. A real-time and accurate object detection technique can help in several critical systems and applications such as self-driving cars (detecting multiple instances of vehicles, humans, etc.), surveillance for public safety, social media intelligence etc. In this talk, we will first understand the problem and go through the complexity levels of different object detection datasets. We will then look at a state-of-the-art of object detection technique : single shot multi-box detection. We will then understand its drawbacks. We will extend this technique using unsupervised learning and feature map confluence to remove the drawbacks. This extension especially helps in those problems and applications where large amount of labeled data is not available. We will then compare performance of these techniques on various datasets. Taking into account sparsity of datasets in various domains, we will then take a deep experimental dive into when and why a particualr object detection technique works and does not work. We will go through a case study to understand different nuiances of different object detection techniques.

Key takeways:

  1. Challenges in doing object detection beyond open source datasets
  2. An end-to-end solution with an open design
  3. A detailed case study and related experiences of applying and comparing different object detection techniques in practice

Note : this work is accepted as a full research publication.

Outline

  1. Challenges in doing object detection beyond open source datasets
  2. An end-to-end solution with an open design
  3. A detailed case study and related experiences of applying and comparing different object detection techniques in practice

Speaker bio

Vijay Gabale is CTO and co-founder of Infilect Technologies. Infilect offers several products in the domain of entertainment, media, and retail by leveraging visual data and providing visual intelligence. Vijay is a PhD from IIT Bombay, and ex-research-scientist IBM Research.

Slides

https://www.dropbox.com/s/8zshjn39p2ulpol/huew-nvidia-gtc-final-2.pdf?dl=0

Comments

Login with Twitter or Google to leave a comment